Conference on Cognitive Computational Neuroscience

Adults, infants, and many animal species can visually perceive the approximate number of items in a set (Feigenson et al., 2004). One theory holds that the human visual system has specialized numerical representations that are independent from other visual features (Dehaene, 2011). However, modeling work indicates that convolutional neural networks trained on object recognition (and untrained networks) have Gaussian feature tuning for numerosity, mirroring neural tuning in primate parietal and frontal cortices (Nasr et al., 2019; Kim et al., 2021). These findings suggest that number representations may emerge from generic architectural and learning constraints. To further test this hypothesis, we examined whether these representations exhibit four signatures of human numerical perception: decreasing number representations for grouped, bounded, and connected items, and increasing number representations for coherently oriented sets. A self-supervised AlexNet model trained on ImageNet exhibited all four signatures, while untrained AlexNet only exhibited humanlike sensitivity to spatial grouping. Thus, humanlike numerical features can emerge when training neural networks to discriminate broadly between different views of the world, without any specialized number-processing constraints.